Skip to content

HunyuanImage 2.1

Model Source Revision Weight license Parity (CPU, fp32)
hunyuanimage-2-1 hunyuanvideo-community/HunyuanImage-2.1-Diffusers 7e7b7a1 Tencent Hunyuan Community License Verified

Catalog: hunyuanimage-2-1 · Models: hunyuanimage-2-1 · Task: t2i

Architecture. A Qwen2.5-VL text model and a ByT5 encoder for quoted text feed a diffusion transformer with dual-stream, single-stream and token-refiner blocks. A 64-channel VAE with 32× compression decodes the latents. The flow-match Euler scheduler uses the two guider configurations from the snapshot, selecting the OCR guider for prompts with quoted text.

Usage.

python -m imagehub.pretrained hunyuanimage-2-1 checkpoints/hunyuanimage-2-1
from imagehub import pipeline

with pipeline("t2i", backend="hunyuanimage", model="checkpoints/hunyuanimage-2-1") as pipe:
    result = pipe('a red beret with the word "Tencent" on it', seed=42)
    result.images[0].save("beret.png")

Settings. steps defaults to 50; width and height default to 2048 and must be multiples of 32. guidance defaults to the selected guider's scale. The other options are seed, latents and negative_prompt.

Memory. The backend loads and releases the text encoders, transformer and VAE in sequence. The ByT5 encoder loads only when the prompt contains quoted text.

Seed compatibility. On CPU, the same seed and settings give the same image as diffusers with generator=torch.Generator("cpu").manual_seed(seed) (validation).

Out of scope. The refiner, distilled variants, LoRA adapters and fp16/bf16 parity are not supported.

License. The weights use the Tencent Hunyuan Community License. imagehub does not distribute weights.